Limits of Ekman Model and Alexithymia

Overview

The study tests whether Ekman’s six “basic” emotions adequately describe complex real-life emotions in text, and examines how alexithymia levels affect people’s emotion labelling.

Background: Models of Emotion

• Categorical models

  • Emotions as discrete, universal “basic” types (e.g. happiness, sadness, anger, fear, disgust, surprise).
  • Ekman’s model defines six basic emotions via cross-cultural agreement on facial expressions.

• Dimensional models

  • Emotions defined along dimensions such as valence, arousal, control.
  • Circumplex models: valence and arousal arranged in a circular space.
  • Vector models: arousal with binary valence (positive vs negative).

• Constructed Emotion theory (Barrett)

  • Emotions are not universal biological reactions identified via facial/physiological markers.
  • Emotions are brain-constructed concepts based on prediction, prior experience, sensory data, and social knowledge.
  • Emotional experience is a simulation of likely bodily responses, not a simple reaction to external events.

• Moral and social emotion perspectives

  • Emotional theory of social psychology: emotions have moral components, support social cohesion, and evolved for group functioning.

• Unnatural and “undefined” emotions in literature

  • “Unnatural emotions”: defamiliarizing, extreme, nonconformist, or logically/physically impossible scenarios.
  • The Dictionary of Obscure Sorrows (DOS) introduces new words for subtle, complex, often unnamed feelings (e.g. “onism,” “sonder”).
  • These emotions are often unfamiliar and poorly captured by standard everyday labels or basic emotion models.

• Affect labelling

  • Putting feelings into words can dampen negative emotions and heighten positive ones.
  • Effectiveness depends on a person’s ability to identify and verbalise emotions.
  • Using specific labels (e.g. amused, joyous) helps emotion awareness and regulation better than generic labels (e.g. happy).
  • Accurate positive emotion labelling improves coping with stress.
  • People who struggle to identify/label emotions tend to experience negative affect more often and more intensely.

• Research gap addressed

  • Question: Can emergent/undefined complex emotions be described by traditional models (such as Ekman’s)?
  • Need to examine whether real-life, context-rich emotional situations map neatly onto basic emotion categories.

Key Concepts: Ekman’s Emotions and Alexithymia

• Ekman’s six basic emotions

  • Happiness, sadness, anger, fear, disgust, surprise.
  • Widely used basis for emotion annotation in AI (e.g. facial expression datasets, text corpora).

• Alexithymia

  • Literally “no words for emotions”.
  • Personality trait involving deficits in identifying, describing, and cognitively processing feelings.
  • Linked to psychiatric conditions (e.g. depression, schizophrenia) and psychosomatic illnesses.
  • Involves both cognitive deficits (recognising, describing, distinguishing emotions) and affective deficits (reduced emotionalising).

• Relevance to emotion labelling

  • Alexithymia is associated with difficulties in verbalising emotions and with less complex emotional language.
  • Investigating alexithymia clarifies how individual differences shape the use of basic vs more nuanced emotion labels.

Research Questions and Aims

• Main research questions

  • Are existing emotion models (especially Ekman’s six basic emotions) sufficient to describe complex, circumstantial emotional situations?
  • Do different levels of alexithymia influence understanding, identification, and perception of complex emotions?

• Aims

  • Test whether Ekman’s six emotions adequately capture emotional experiences from two text corpora.
  • Examine human agreement when using Ekman labels on:
    • A dataset already annotated with Ekman emotions (ELTEA17 tweets).
    • A literary dataset intended to express “undefined” emotions (DOS entries).
  • Assess how alexithymia level affects the tendency to use basic emotions vs “other” labels.

Participants and Instruments

Participants

• 114 adults (20 males, 94 females).
• Age groups roughly split into:
  • Generation X: 42 years and older (53.5%).
  • Generation Z: 41 years and younger (46.5%).
• Informed consent obtained; study approved by ethics committee; conducted according to the Declaration of Helsinki.

Perth Alexithymia Questionnaire (PAQ)

• Purpose

  • Measure alexithymia across multiple dimensions in healthy adults.

• Main components

  • Difficulty identifying feelings (DIF).
  • Difficulty describing feelings (DDF).
  • Externally oriented thinking (EOT) – tendency to focus on external events rather than internal emotions.

• Structure

  • 24 items, 10 subscales:
    • Negative-Difficulty identifying feelings (N-DIF).
    • Positive-Difficulty identifying feelings (P-DIF).
    • Negative-Difficulty describing feelings (N-DDF).
    • Positive-Difficulty describing feelings (P-DDF).
    • General-Externally oriented thinking (G-EOT).
    • General-Difficulty identifying feelings (G-DIF).
    • General-Difficulty describing feelings (G-DDF).
    • Negative-Difficulty appraising feelings (N-DAF).
    • Positive-Difficulty appraising feelings (P-DAF).
    • General-Difficulty appraising feelings (G-DAF).
  • Response format: 1 (strongly disagree) to 7 (strongly agree) Likert scale.

• Translation

  • English PAQ independently translated into Italian by the authors; final items agreed by consensus due to lack of existing Italian version.

Emotion Annotation Task

• 20 sentences describing emotional experiences.

  • 10 sentences from The Dictionary of Obscure Sorrows (DOS).
  • 10 sentences from ELTEA17 tweets annotated with Ekman’s emotions.

• Annotation options per sentence

  • One of Ekman’s six emotions.
  • “Other”
    • Participants could specify a non-Ekman emotion word and/or multiple emotions.
  • “No emotion”
    • Chosen when participants perceived no emotion in the sentence.

• Selection of sentences

  • ELTEA17 tweets chosen to resemble DOS entries:
    • Describe a specific situation.
    • Include temporal context (e.g. “when you feel that…”).
    • Show complexity/ambiguity in emotional narration or temporal/physical aspects.
    • Avoid overly personal/specific references (e.g. specific relatives or locations).
  • DOS entries chosen to cover both pleasant (cheerful) and unpleasant (sad) emotional tones.
  • All texts translated into Italian.

Procedure

• Data collection

  • Online questionnaires built with Google Forms.
  • Distributed through social media (Facebook, WhatsApp, Instagram, etc.).
  • Aimed to reduce interviewer influence and reach a large, diverse adult sample.

• Questionnaire structure

  • Page 1: Consent form, demographic data (age, gender).
  • Page 2: PAQ items (alexithymia assessment).
  • Page 3: ELTEA sentences emotion annotation.
  • Page 4: DOS sentences emotion annotation.

• Implementation details

  • Approximate completion time: 10 minutes.
  • Only closed questions; all items mandatory.
  • Multiple submissions disabled to avoid duplication.
  • All participants completed the full questionnaire.

Grouping by Alexithymia Level

• Based on total PAQ score:

  • Group 1 (Low alexithymia): 1 standard deviation or more below mean (N = 19).
  • Group 2 (Medium alexithymia): less than 1 standard deviation from mean (N = 81).
  • Group 3 (High alexithymia): 1 standard deviation or more above mean (N = 14).

• Reliability

  • Total PAQ scale mean: 80.38; SD: 27.52.
  • Cronbach’s alpha: 0.94 (high internal consistency).

Summary Table: Design and Measures

Results: Agreement in Emotion Labelling

Overall group differences

• One-way ANOVA on agreement between groups
  • No significant differences between alexithymia groups in overall inter-group agreement:
    • ELTEA: F(2,26) = 0.649, p = 0.508, η² = 0.051.
    • DOS: F(2,25) = 0.715, p = 0.499, η² = 0.054.
• However, within-group agreement patterns differed across alexithymia levels.

Agreement within Alexithymia Groups

• Low alexithymia (Group 1)

  • ELTEA: agreement with original annotations in 2 out of 10 items.
  • Only one ELTEA item with >50% agreement: ELTEA3 (“Happiness”).
  • DOS: one item (DOS6, “Happiness”) with >50% agreement.
  • Agreement for other items typically between about 21.1% and 47.40%.

• Medium alexithymia (Group 2)

  • ELTEA: better match with original labels; disagreement in only 4 of 10 items.
  • For 5 of the 6 agreeing items, agreement ranges ~14.8–39.5%;
  • Highest agreement again on ELTEA3 (“Happiness”) with 77.8% agreement.
  • DOS: 7 of 10 items have agreement around 50%; remaining items show agreement between ~17.3% and 37%.

• High alexithymia (Group 3)

  • ELTEA: agreement in 6 of 10 items; agreement ranges from 28.6% to 92.9%.
  • ELTEA3 (“Happiness”) again shows the highest agreement (92.9%).
  • DOS: equal or greater than 50% agreement in half of the items.

• Consistent disagreement across groups

  • For ELTEA5, ELTEA6, and ELTEA10, no group matches the original annotation well.

• Interpretation

  • High agreement tends to concentrate in clearly positive, simple situations (e.g. ELTEA3, DOS6 labelled as happiness).
  • Low levels of agreement are widespread, suggesting that complex, circumstantial emotions resist neat Ekman categorisation.
  • Variation in which items produce low agreement across groups points to limitations of the model rather than a few “bad” annotations.

Results: Use of Ekman Emotions vs “Other” by Alexithymia Level

Statistical assumptions

• Shapiro–Wilk tests show non-normal distributions for counts of each emotion.
• Non-parametric tests (Kruskal–Wallis, Bonferroni-corrected post-hoc) applied.

Overall frequency of emotional labels

• Significant overall group difference in total use of affective labels (canonical emotions).

  • H = 11.99, p = 0.002.

• Post-hoc comparisons (Bonferroni corrected, p < 0.016 threshold):

  • High vs low alexithymia: high group uses emotional labels more often (M = 77.78 vs 40.21; p = 0.003).
  • No significant differences:
    • Low vs medium (M = 58.04; p = 0.102).
    • Medium vs high (p = 0.117).

Overall frequency of “other” (non-canonical) labels

• Significant difference in use of “other” category (non-canonical affect labels).

  • H = 11.99, p = 0.002.

• Post-hoc comparisons:

  • Low vs high alexithymia: low group uses “other” more often (M = 74.68 vs 37.21; p = 0.003).
  • No significant differences:
    • Low vs medium (M = 56.97; p = 0.106).
    • Medium vs high (p = 0.116).

• Interpretation

  • Individuals with higher alexithymia rely more on traditional Ekman categories to interpret scenarios.
  • Individuals with lower alexithymia more frequently consider scenarios as non-canonical or requiring novel/other labels.

Emotion Frequency and Alexithymia Subscales

Subscale N-DIF (Difficulty Identifying Negative Feelings)

• Kruskal–Wallis: significant difference for use of fear.
  • H = 8.832, p = 0.012.
• Post-hoc
  • High vs low alexithymia: high group more often labels scenarios as fear (M = 67.03 vs 32.00; p = 0.011).
  • No significant differences:
    • Low vs medium (M = 57.21; p = 0.070).
    • Medium vs high (p = 0.522).

Subscale N-DDF (Difficulty Describing Negative Feelings)

• Use of canonical affective categories overall

  • H = 10.41, p = 0.005.
  • High vs low: high group uses canonical labels more (M = 72.94 vs 40.71; p = 0.007).
  • No differences: low vs medium (M = 57.83; p = 0.130); medium vs high (p = 0.224).

• Use of “other” labels

  • Low vs high: low group uses “other” more (M = 73.84 vs 42.13; p = 0.009).
  • No differences: low vs medium (M = 57.25; p = 0.151); medium vs high (p = 0.223).

Subscale G-EOT (General Externally Oriented Thinking)

• Use of canonical affective categories

  • H = 12.03, p = 0.002.
  • High vs low: high group uses canonical labels more (M = 78.81 vs 43.88; p = 0.021).
  • Low vs medium: medium group more than low (M = 61.81; p = 0.025).
  • Medium vs high: no difference (p = 0.503).

• Use of “other” labels

  • Low vs high: low group uses “other” more (M = 71.186 vs 36.18; p = 0.020).
  • Low vs medium: low group uses “other” more than medium (M = 53.15; p = 0.024).
  • Medium vs high: no difference (p = 0.506).

Subscales G-DDF (General Difficulty Describing Feelings) and G-DAF (General Difficulty Appraising Feelings)

• General pattern

  • High alexithymia scores linked to less frequent use of “other” labels compared with low alexithymia.

• G-DDF (subscale 7)

  • H = 6.86, p = 0.03.
  • High vs low: high group uses “other” less (M = 45.14 vs 71.05; p = 0.043).
  • No differences: low vs medium (M = 57.70; p = 0.371); medium vs high (p = 0.370).

• G-DAF (subscale 8)

  • H = 8.62, p = 0.01.
  • High vs low: high group uses “other” less (M = 42.50 vs 74.15; p = 0.022).
  • No differences: low vs medium (M = 58.53; p = 0.340); medium vs high (p = 0.156).

Pearson Correlations

• Fear and alexithymia

  • Significant positive association between use of fear labels and higher alexithymia scores on:
    • N-DIF (Negative-Difficulty identifying feelings).
    • G-DIF (General-Difficulty identifying feelings).
  • p = 0.01 for fear–alexithymia correlation.

• Generic use of affective terminology increases with scores on several negative-feeling subscales:

  • N-DDF (Difficulty describing negative feelings): p = 0.001.
  • G-EOT (General externally oriented thinking): p < 0.001.
  • G-DIF: p = 0.01.
  • G-DDF: p = 0.007.
  • G-DAF: p = 0.003.

• Interpretation

  • Higher difficulty in identifying, describing, and appraising feelings corresponds to greater reliance on standard emotion labels, especially fear and other negative emotions.

Chi-Square Analyses for Specific Items

• ELTEA dataset

  • Significant difference in preference for fear on item 4.
  • χ² = 10.99, p = 0.004.
  • Group 3 (high alexithymia) shows especially high annotation rate of fear for this item.

• DOS dataset

  • Significant difference for item 5 regarding use of “other” category.
  • χ² = 44.98, p < 0.001.
  • Group 3 (high alexithymia) shows lower inclination to use “other” labels compared with groups 1 and 2.

Summary Table: Main Findings by Alexithymia Level

Discussion: Adequacy of Ekman’s Model

• Low inter-subject agreement

  • For both ELTEA17 (Ekman-annotated tweets) and DOS (undefined emotions), overall agreement among subjects is low.
  • Agreement is particularly low for participants with low alexithymia who show greater diversity in chosen labels.

• Limited explanatory power of basic emotions

  • If poor agreement was limited to a few items, one could blame faulty original annotations.
  • Instead, disagreement is widespread and differs by group, indicating that Ekman’s six emotions often fail to capture complex, situationally rich experiences.

• Alignment with prior critiques

  • Supports arguments that models based solely on basic emotions (like Ekman’s) are too rigid for real-life emotional diversity.
  • Previous AI work using Ekman’s categories to train emotion recognition systems may overlook complex or mixed emotions.

• Implications for emotion modelling and AI

  • Real-life emotional situations involve nuance, context, and “undefined” emotions that cannot be reduced to a small fixed set.
  • There is a need to extend or integrate existing models to capture richer emotional structures in text and multimodal data.

Discussion: Alexithymia and Emotion Labelling

• High alexithymia: more basic labels, less nuance

  • Individuals with high alexithymia rely heavily on traditional Ekman categories, especially for negative emotions like fear.
  • They use non-canonical “other” labels significantly less than low-alexithymia individuals.

• Low alexithymia: higher granularity and flexibility

  • Individuals with low alexithymia show more diverse and specific emotional labelling.
  • They more often judge emotional scenarios as not well captured by canonical labels and choose “other” or multi-label responses.

• Relation to prior findings

  • Consistent with evidence that alexithymic individuals:
    • Struggle to recognise facial emotional expressions.
    • Have impaired emotional linguistic processing.
    • Use less complex emotional vocabulary.

• Coding vs judging processes

  • Coding: automatic detection of emotional cues.
  • Judging: cognitive evaluation and interpretation of emotional stimuli.
  • Alexithymic individuals may detect basic emotional cues but have difficulty evaluating and fine-graining them.

• Bias toward negative emotions, especially fear

  • Alexithymic participants show a notable tendency to assign fear in ambiguous situations.
  • Links to difficulties mapping negative emotions (like fear) into sensorial/embodied representations.

• Possible neural basis

  • fMRI studies show relationships between alexithymia and altered activity in the anterior cingulate cortex (rostral anterior and posterior regions).
  • Suggest reduced efficiency in the interplay between affective and cognitive processes.

• Psychological implications

  • Greater granularity in emotion labelling (seen in low alexithymia) supports better emotion regulation and resilience.
  • Conversely, reliance on coarse basic categories might limit emotional understanding and regulation, reinforcing negative biases.

Computational and Ontological Developments

• Motivation from AI research

  • Many computational systems use datasets annotated only with Ekman’s basic emotions.
  • This study reveals these models are insufficient for realistic, complex emotional situations.

• SPICE project context

  • Focuses on cultural engagement, social inclusion, and empathy development.
  • Utilises knowledge graph infrastructure and multiple emotion theories in an ontology network.

• Knowledge-graph approach

  • Emotion-oriented resources (datasets, lexicons) represented as interoperable knowledge graphs with formal semantics.
  • Example: Framester, a wide-coverage linguistic linked data hub based on cognitive frame semantics.

• New modelling efforts

  • Development of the Emotion Frame Ontology:
    • Integrates multiple emotion theories.
    • Provides a first-order theory representing both real-world situations and emotional states.
    • Aims to abstract over existing models while staying flexible enough for complex cases.
  • Use of neuro-symbolic systems:
    • Combine explainable graph structures and automated reasoning with learnable emotional patterns from multimodal data.

• Goal

  • Build an extended, flexible computational model for emotion situations better aligned with human, real-world emotional experiences.

Limitations

• Gender imbalance

  • Sample heavily skewed toward females (94 females vs 20 males).
  • Limits generalisability and may influence patterns of emotion labelling and alexithymia.

• Nature of corpora

  • Tweets (ELTEA17) are brief and constrained, potentially limiting expression of nuanced emotions.
  • Future work could use longer, more naturalistic corpora annotated with Ekman emotions to test if inter-rater agreement improves.

• Online survey bias

  • Self-selection bias: participants opting into an online questionnaire may not represent the general population.
  • Social desirability bias: participants might choose labels they think are socially acceptable.

Key Terms & Definitions

• Ekman’s basic emotions

  • Six supposedly universal, biologically based emotions: happiness, sadness, anger, fear, disgust, surprise.

• Categorical emotional models

  • Theories that classify emotions into discrete types or categories (e.g. basic emotions).

• Dimensional emotional models

  • Theories that represent emotions as positions in a continuous space (e.g. valence, arousal, control).

• Constructed Emotion theory

  • Approach positing that emotions are brain-constructed concepts based on predictions and social learning, not innate universal states.

• Alexithymia

  • Personality trait characterised by difficulties identifying, describing, and mentally representing emotions, and by externally oriented thinking.

• Affect labelling

  • Process of verbally naming or describing one’s emotional state, often linked to regulation and awareness of emotions.

• Complex/circumstantial/undefined emotions

  • Emotional experiences tightly tied to specific contexts, often not captured by standard everyday emotion words or traditional models.

• ELTEA17 dataset

  • Entity-Level Tweets Emotional Analysis dataset annotated according to Ekman’s six emotions.

• The Dictionary of Obscure Sorrows (DOS)

  • Literary work introducing invented terms for subtle, complex, and often unnamed emotional experiences.

• Knowledge graph

  • Structured representation of entities and relations using formal semantics, allowing interoperability and automated reasoning.

Action Items / Next Steps

• For students and researchers

  • Review how alexithymia may bias emotion research that relies solely on basic labels.
  • Consider using richer emotion models or mixed categorical–dimensional approaches in experimental designs.

• For computational modelling

  • Incorporate more flexible ontologies (e.g. Emotion Frame Ontology) when building emotion-aware AI systems.
  • Extend datasets beyond Ekman’s six emotions to include complex and context-dependent emotional descriptions.

• For future empirical work

  • Replicate the study with more balanced gender and broader demographics.
  • Use longer, narrative-based corpora annotated with multiple emotion models.
  • Further test the relationship between emotional granularity, alexithymia, and psychological outcomes such as resilience.